memory mechanisms in man & machine · •two cases of free identification • 90 year old woman...

Memory Mechanisms in Man & Machine

Simon ThorpeDirector of the Brain and Cognition Research Centre (CerCo)Director of the Toulouse Mind & Brain Institute (TMBI)BrainChip IncToulouse, France

The M4 Project : ERC Grant (2013-19)

1. Humans can recognise visual and auditory stimuli that they have not experienced for decades

• Two cases of free identification• 90 year old woman who

said “Balsac” for “Un amour de Balsac”

• 77 year old man who said “Camember” for “Les faciétés du sapeur Camember• 70 TV themes never rebroadcast

for several decades

1. Humans can recognise visual and auditory stimuli that they have not experienced for decades

2. Recognition after very long delays is possible without ever reactivating the memory trace in the intervening period

• Testing after a month

• Impossible to rehearse

• Looped sounds also work

• It’s not just the start

• Scrambled sounds also work

• The learning must work on tiny snippets only 10-20 ms long

3. These very long term memories require an initial memorisation phase, during which memory strength increases roughly linearly with the number of presentations

• Frame Rate : 2, 4, 6, 9, 10, 12, 15, 20, 30, 60…

• Repetitions : 1, 2, 3, 4, 5, 6, 7, 8…

• Distractors between targets : 1, 2, 3, 4, 5….

• 556 different conditions!

BrainSpotting

3. These very long term memories require an initial memorisation phase, during which memory strength increases roughly linearly with the number of presentations

• Roughly linear increase from 1 to 5 repeats

• But - only a short interval between training and testing

• Would this also apply for Long Term Memories?

3. These very long term memories require an initial memorisation phase, during which memory strength increases roughly linearly with the number of presentations

• Thunell & Thorpe (under review)

• Performance doesn’t decrease with testing delay!

• At least for a few minutes….

4. A few tens of presentations can be enough to form a memory that can last a lifetime

• Images from the DMS-48 test• 243 subjects tested between 2002 and 2006• 63 tested in 2016

6. Storing such very long-term memories involves the creation of highly selective “Grandmother Cells” that only fire if the original training stimulus is experienced again

• Still just a theoretical hypothesis

• Single unit recording from cortical neurons in patients could be the ultimate test

7. The neocortex contains large numbers of totally silent cells (“Neocortical Dark Matter”) that constitute the long-term memory store

8. Grandmother Cells can be produced using simple spiking neural network models with Spike- Time Dependent Plasticity (STDP) and competitive inhibitory lateral connections

9. This selectivity only requires binary synaptic weights that are either “on” or “off”, greatly simplifying the problem of maintaining the memory over long periods

•A new highly efficient algorithm

•JAST

•Jake, Amir, Simon & Tim!

•System with Binary weights

•Input neurons fire spikes asynchronously

•Output neurons function as synchrony detectors

•Learning rule adjusts the weights to match incoming spike patterns

JAST

10. Artificial systems using memristor-like devices can implement the same principles, allowing the development of powerful new processing architectures that could replace conventional computing hardware

March202017 Sep102017

10. Artificial systems using memristor-like devices can implement the same principles, allowing the development of powerful new processing architectures that could replace conventional computing hardware

AKIDA

• 1.2 million neurons

• 10 billion connections

• 7 mm chip

• On-chip learning using JAST

• $10 fabrication cost

The M4 Project : ERC Grant (2013-19)

Why is this important? • The Deep Learning Revolution

Human performance

Superhuman performance

Supervision

Supervision

• Reinforcement Based Learning

• Deep Mind’s Chess system learned from scratch in 4 hours

Deep Learning Chips

• Gyrfalcon Technology Inc

• Movidius

• Plus

• Graphcore

• Anotherbrain.ai

• GrAI Matter Labs

• Google’s Edge TPU

• Plus• Apple• Microsoft• NVidia• Amazon• ARM• IBM• ….

Natural Intelligence

•The implications of the M4 project

• Humans don’t learn like Deep Learning Systems

• We don’t need hundreds of millions of training cycles with labelled data

• Humans learn in 2-5 repeats

• Unsupervised learning

• We now have learning rules that can do the same thing with relatively simple spiking neural networks

• Implementation on inexpensive custom chips

• Grandmother Cells and Neocortical Dark Matter are very important for power consumption

Implications for society• Very powerful Deep Learning chips for $10 could replace many paid jobs

• AKIDA style intelligent chips are potentially an even greater threat!

• The resulting technological unemployment could prove catastrophic

• My proposed solution

• Universal Basic Income!

• Calum Chace • Economic Singularity Club Think Tank

• Forthcoming book

• “Stories from 2045”

• 24 positive stories

• 13 negative ones

Questions or Comments?